pipecat-akapulu

Pipecat processor for an Akapulu Labs talking avatar. Maintained by Akapulu Labs.

AkapuluAvatar sits after your speech model. It sends assistant audio to the Akapulu worker. The worker joins the Daily room already stored on the session and publishes the face and the played audio. This processor does not publish either.

Tested with Pipecat 1.12.0.

Install

uv add pipecat-akapulu

Use

Create the Daily room yourself, then POST /api/live-sessions/ from your server with your Akapulu API key before the bot starts. The body is avatar_id, show_ai_label, room_url, and avatar_token. The response is session_id.

from pipecat_akapulu import AkapuluAvatar

avatar = AkapuluAvatar(sample_rate=24000, session_id=session_id)

sample_rate is a positive multiple of 25 and must match the assistant audio frames. tail_secs defaults to 0.6, which is how long assistant audio stays open after speech stops.

Leave transport audio out and video out off. The worker is the Daily publisher for the face and the played audio.

AKAPULU_API_URL selects the API origin. Unset, the processor uses https://akapulu.com. A set value that is empty or not an http or https origin raises.

On StartFrame the processor posts {} to POST /api/live-sessions/<session_id>/status/. It polls GET on that same route for the worker socket and for loading status, and pushes each status body as an RTVI server message.

Example

examples/foundational.py is a Daily bot: speech-to-text, a model, speech output, then AkapuluAvatar. It greets when the Daily client connects.

uv add "pipecat-ai[daily,deepgram,cartesia,google,silero]==1.12.0" pipecat-akapulu

Set DAILY_ROOM_URL, DAILY_TOKEN, and AKAPULU_SESSION_ID from the session you created. Also set DEEPGRAM_API_KEY, GOOGLE_API_KEY, and CARTESIA_API_KEY. Then:

python examples/foundational.py